prompt-injection-defense 开源项目推荐（205 个）

requie /LLMSecurityGuide

大型语言模型（LLMs）安全保护的综合参考资料。涵盖OWASP生成式AI十大风险、即时注入、对抗性攻击、现实世界事件及实用防御。包含红队工具、防护措施和缓解策略目录，帮助开发者、研究人员和安全团队负责任地部署人工智能。

StackOneHQ /defender

Open source prompt injection protection for Agents calling tools (via MCP, CLI or direct function calling). Detect and defend against prompt injection attacks. 22MB, CPU-only, < 10ms latency.

R3dShad0w7 /PromptMe

PromptMe is an educational project that showcases security vulnerabilities in large language models (LLMs) and their web integrations. It includes 10 hands-on challenges inspired by the OWASP LLM Top 10, demonstrating how these vulnerabilities can be discovered and exploited in real-world scenarios.

FIND-Lab /AgentWard

AgentWard – Built for all, hardened for OpenClaw.

occludra /gateway

Self-hosted AI security proxy. Redact PII, block prompt injection, route to any LLM provider. OpenAI-compatible.

2alf /Heimdall

Tamper-evident integrity monitor for the MCP config & server files your local AI agents load.

metacognitionai /zeos

ZEOS: a transformer operating system — priority-preemptive scheduling, protected mode, and deterministic replay for LLM jobs

cyb3rmik3 /prompt-injection-email-samples

Open test set of .eml emails for checking how email security controls and AI mailbox assistants handle indirect prompt injection. It crosses three intents (system-prompt disclosure, data exfiltration, tool discovery) with three delivery methods (plaintext, hidden HTML, Base64), plus a benign control. All content is fictional and non-routable.

sleeepeer /PISanitizer

[EMNLP 2026] PISanitizer: Preventing Prompt Injection to Long-Context LLMs via Prompt Sanitization

webpro255 /agentlock

An adversarially benchmarked reference implementation for pre-action AI agent authorization. Provenance-based gating for LLM agent tool calls: deny-by-default permissions, parameter lineage, signed receipts, audit logging.

Top-Celestial-Company-Ltd /DROS-VajraClaw-Hacker

⚡ DROS™ VajraClaw (Hacker Edition): Free Multi-Agent Runtime Security & Docker Governance Gateway with W3C DID & <1μs Enforcement

dronefreak /PromptScreen

Protect your LLMs from prompt injection and jailbreak attacks. Easy-to-use Python package with multiple detection methods, CLI tool, and FastAPI integration.

trickymind1324 /personal-finance-agent

Personal finance agent that decides whether an expense is safe — pay in full, split, use installments, or wait — from a 90-day balance forecast over real transaction history, with a deterministic decision core and a narrow multimodal evidence layer for receipts and messages.

OMD-123 /browser-verified-skill

Browser automation with evidence. An Agent Skill that lets an AI agent drive a real browser while every action passes a policy gate, every run leaves an auditable artifact, and the run compiles into a plain Playwright script that replays for free.

SECBLOK /belay

Belay is an open-source, local-first security layer for AI coding agents (Claude Code, Codex, Cursor, OpenClaw, Hermes Agent and MCP) that blocks dangerous commands, secret leaks, and prompt injection at the tool-call boundary in under 100ms — no LLM in the decision path by default, no cloud, no phone-home.

intentframe /agent-integrations

IntentFrame security plugin for Hermes Agent (Nous Research) — an external policy checkpoint that gates terminal, code, file, and cron tool calls before they run on your machine.

butterclaw-tech /butterclaw

Runtime security enforcement for autonomous AI agents.

kevinl05 /promptmenot

Detect and sanitize prompt injection attacks in Rails apps. Protects against direct injection (users hacking your LLMs via form inputs) and indirect injection (malicious prompts stored for other LLMs to scrape). ~70 detection patterns across 7 attack categories with configurable sensitivity levels. Now includes resource extraction detection pattern

ilyautov /cordon

A prompt-injection firewall for AI agents, with no AI inside: your agent reads anything and takes orders only from you. For Claude Code, Gemini CLI, MCP hosts and LangChain. | Файрвол от промпт-инъекций для ИИ-агентов без ИИ внутри: агент читает что угодно и слушается только вас.

FlapPearLabs /zhihu-grabber-toolkit

ZhiHu CLI PRO — Cross-question deep research for Zhihu & AI Agents. Search, capture, verify, analyze large corpora, preserve evidence lineage, and synthesize across questions.